[Mesa-dev] [PATCH 5/9] i965/nir: Emit MUL + ADD for MAD on GEN <= 5

[Jason Ekstrand]

On Tue, Mar 17, 2015 at 7:35 PM, Connor Abbott <cwabbott0 at gmail.com> wrote:
> Instead of doing this, I think it would be better to add something to
> nir_shader_compiler_options and do it in nir_opt_algebraic, similar to
> what Eric has already done for other things. I've seen some shaders
> where we transform a mul + a series of adds into a series of mad's,
> which is good on gen6+ (at the very least, we reduce sched
> dependencies, and we may be able to eliminate the mul) but if we do
> that and then just naively lower it back we'll get worse results than
> we started with. Similar thing with the previous lrp patch.

Yes, that may be better in some cases.  I'm still not sure I like the
whole pile of booleans thing...

> On Tue, Mar 17, 2015 at 10:17 PM, Jason Ekstrand <jason at jlekstrand.net> wrote:
>> ---
>>  src/mesa/drivers/dri/i965/brw_fs_nir.cpp | 8 +++++++-
>>  1 file changed, 7 insertions(+), 1 deletion(-)
>>
>> diff --git a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
>> index 5da8423..41f9ae2 100644
>> --- a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
>> +++ b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
>> @@ -1234,7 +1234,13 @@ fs_visitor::nir_emit_alu(nir_alu_instr *instr)
>>        break;
>>
>>     case nir_op_ffma:
>> -      inst = emit(MAD(result, op[2], op[1], op[0]));
>> +      if (brw->gen >= 6) {
>> +         inst = emit(MAD(result, op[2], op[1], op[0]));
>> +      } else {
>> +         fs_reg temp = vgrf(glsl_type::float_type);
>> +         emit(MUL(temp, op[0], op[1]));
>> +         inst = emit(ADD(result, temp, op[2]));
>> +      }
>>        inst->saturate = instr->dest.saturate;
>>        break;
>>
>> --
>> 2.3.2
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> http://lists.freedesktop.org/mailman/listinfo/mesa-dev